Hierarchical Bayesian Strata Targeting

Picture this

Imagine rating dishes at different branches of a restaurant chain. If a new branch has served only two customers, judging that location purely on those two reviews is unreliable, but assuming it behaves identically to every other branch is also inaccurate. Combining the chain-wide average with local feedback from that specific branch creates a balanced estimate; as more local reviews accumulate, the model relies progressively more on local data and less on the global average.

What the evidence says

The optimal targeted treatment policy produced an estimated 1.7 percentage point increase in 6-week employment over control (a 35% gain, 95% credible set [0.001, 0.034]), compared to a 0.5 percentage point gain for the optimal non-targeted policy.

Who was studied

N = 3,770 jobseekers (1,663 Syrian refugees and 2,107 Jordanians) divided into 16 distinct demographic strata defined by nationality, gender, education level, and prior wage work experience.

How

Hierarchical Bayesian Beta-Bernoulli model with Markov Chain Monte Carlo sampling (1,000 burn-in iterations, 10,000 replications) estimating stratum-specific success parameters as weighted averages of local stratum outcomes and population-wide treatment outcomes.

What to do

Estimate heterogeneous treatment effect parameters using a hierarchical Bayesian model that reweights local stratum success rates against shared hyper-parameters to guide targeted allocation.

From the source

"At each time period t, the treatment effect of each treatment d in each stratum x is estimated as a weighted average of the observed success rate for d in x and the observed success rates for d across all other strata."

An_Adaptive_Targeted_Field_Experiment_Job_Search_Assistance_for.pdf

Tags

• hierarchical bayesian model
• heterogeneous treatment effects
• targeting